Chinese AI startup DeepSeek plans to develop a specialized chip optimized for inference tasks to reduce dependence on Nvidia and Huawei and optimize the economics of scaling its models.
What Happened
According to sources, DeepSeek is moving toward vertical integration by developing its own solution for executing user queries. Unlike universal and expensive chips used for model training, this new project will target scalable and less energy-intensive execution of inference tasks.
Context
Developing proprietary hardware is a strategic move amid US export restrictions on advanced semiconductor manufacturing technologies. DeepSeek, which has already established itself with efficient open models, aims to ensure stable availability of computing power.
Why It Matters for the Industry
Specializing in inference could change the economics of scaling AI services, allowing for reduced requirements for expensive HBM (High Bandwidth Memory) in favor of more affordable architectural solutions. This reinforces the global trend of companies vertically integrating from model development to creating their own hardware.
Why It Matters for Users
Successful implementation of the project could significantly lower the cost and increase the speed of large AI agents and applications. This could lead to changes in cost-per-token and make the transition from expensive APIs to self-hosted solutions on specialized hardware more profitable.
What Is Not Yet Known / Limitations
The project is in the development stage according to sources; there is currently no practical application in production.
Sources
Author
Look at AI, Editorial Team